Financial data set determination method and system and server

Abstract

The invention provides a financial data set determination method and system and a server, relates to the technical field of data processing, and can integrate a graph neural network into a generator and a discriminator of a generative adversarial network. According to the method, a virtual fraud picture sample with a complex topological structure, multi-entity association and dynamic evolution characteristics can be directly learned and generated on picture structure data, so that a high-quality financial data set is obtained for training a financial fraud detection model, and the performance of the financial fraud detection model is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a financial data set determination method, system and server. BACKGROUND

[0002] In the fields of finance, insurance, enterprise internal control, etc., using machine learning models to automatically detect financial fraud behaviors (such as false reimbursement, credit card fraud, money laundering, insurance fraud, etc.) has become the main implementation means at present. These fraud detection models are usually supervised learning classifiers, which need a large amount of labeled data for training, including a large number of normal transaction samples and sufficient and diverse fraud transaction samples. Only by learning enough fraud patterns can the model accurately identify new and unknown fraud behaviors in actual application.

[0003] When building a fraud detection model, the core problem faced is the high imbalance of data. In real-world data sets, the number of normal transaction samples is often thousands or even tens of thousands of times that of fraud samples, and fraud samples are not only few in number but also single in pattern. Traditional oversampling methods generate new samples by linear interpolation between existing minority class samples, but this method is prone to generate ambiguous and meaningless samples, and cannot create truly novel fraud patterns, which is of limited help to improve the generalization ability of the model. SUMMARY

[0004] The purpose of the present application is to provide a financial data set determination method, system and server, which can integrate graph neural networks into the generator and discriminator of the generative adversarial network, can directly learn and generate virtual fraud graph samples with complex topological structure, multi-entity association and dynamic evolution characteristics on graph structure data, so as to obtain high-quality financial data sets for training of financial fraud detection models, thereby significantly improving the performance of financial fraud detection models.

[0005] In a first aspect, the embodiments of the present application provide a financial data set determination method, which is used in the training process of a financial fraud detection model, and the method comprises: obtaining a financial transaction graph corresponding to original financial transaction data, and constructing a financial fraud knowledge graph using heterogeneous multi-modal features in the original financial transaction data; determining a virtual fraud transaction graph corresponding to the original financial transaction data according to the financial transaction graph and the financial fraud knowledge graph; constructing a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph, and constructing a generative adversarial network corresponding to the original financial transaction data using the generator and the discriminator; The graph feature data corresponding to the financial transaction graph and the condition information data corresponding to the financial fraud knowledge graph are determined, the generator in the generative adversarial network is used to determine the financial fraud sample graph corresponding to the graph feature data and the condition information data, and after the graph structure and attribute value of the financial fraud sample graph are evaluated according to a preset evaluation index by the discriminator in the generative adversarial network, the financial fraud sample graph satisfying the evaluation index is obtained. The financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph are obtained based on the financial transaction graph, and the financial data set used for training the financial fraud detection model is constructed by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

[0006] Optionally, the step of obtaining the financial transaction graph corresponding to the original financial transaction data and constructing the financial fraud knowledge graph by using the heterogeneous multi-modal features in the original financial transaction data comprises: The data source corresponding to the financial data set is obtained, and the original financial transaction data is determined by using the enterprise transaction data, the bank flow data, the accounting voucher data, the enterprise business data, the judicial judgment data and the media report data determined by the data source; After the original financial transaction data is standardized and encoded according to the data features and category features corresponding to the original financial transaction data, the heterogeneous feature data corresponding to the original financial transaction data is obtained; The specific entity data and the specific relationship data corresponding to the original financial transaction data are determined by using the heterogeneous feature data, the nodes corresponding to the financial transaction graph are determined according to the specific entity data, and the edges corresponding to the financial transaction graph are determined according to the specific relationship data, and then the financial transaction graph is constructed by using the nodes and the edges; The key entity data, the key relationship data and the financial fraud mode data corresponding to the original financial transaction data are determined by using the heterogeneous feature data, the embedding representation vector corresponding to the financial fraud knowledge graph is determined by using the key entity data, the key relationship data and the financial fraud mode data, and the financial fraud knowledge graph is constructed according to the semantic information and the structural information corresponding to the embedding representation vector.

[0007] Optionally, the virtual fraud transaction graph corresponding to the original financial transaction data is determined according to the financial transaction graph and the financial fraud knowledge graph, comprising: The graph neural network corresponding to the original financial transaction data is constructed by using the financial transaction graph and the financial fraud knowledge graph; The time sequence dependence parameter, the correlation parameter, the heterogeneous node feature parameter and the heterogeneous edge feature parameter corresponding to the financial transaction graph and the financial fraud knowledge graph are obtained based on the graph neural network; The virtual fraud transaction graph is generated according to the time sequence dependence parameter, the correlation parameter, the heterogeneous node feature parameter and the heterogeneous edge feature parameter.

[0008] Optionally, the generator and the discriminator corresponding to the original financial transaction data are constructed based on the virtual fraud transaction graph, including: The financial fraud knowledge graph corresponding to the graph neural network is determined based on the virtual fraud transaction graph, and the first input data and the first output data corresponding to the financial fraud knowledge graph are obtained; The financial transaction graph corresponding to the graph neural network is determined based on the virtual fraud transaction graph, and the second input data and the second output data corresponding to the financial transaction graph are obtained; The generator is constructed using the first input data and the first output data, and the discriminator is constructed using the second input data and the second output data.

[0009] Optionally, the generative adversarial network corresponding to the original financial transaction data is constructed using the generator and the discriminator, including: The virtual fraud graph corresponding to the original financial transaction data is generated using the generator; The real fraud graph contained in the original financial transaction data is obtained, and the discriminator is used to obtain the discrimination result of the real fraud graph and the virtual fraud graph; The generator and the discriminator are alternately trained using the WGAN-GP loss function or the Wasserstein distance loss function, and the discrimination result is obtained in real time; When the discrimination result meets the preset threshold condition, the generative adversarial network is constructed based on the current generator and the discriminator.

[0010] Optionally, the graph feature data corresponding to the financial transaction graph and the condition information data corresponding to the financial fraud knowledge graph are determined, and the financial fraud sample graph corresponding to the graph feature data and the condition information data is determined using the generator in the generative adversarial network, including: The extraction strategy corresponding to the financial fraud sample graph is determined based on the original financial transaction data; The graph feature data is extracted from the financial transaction graph using the extraction strategy, and the condition information data is extracted from the financial fraud knowledge graph using the extraction strategy; Random noise is set using the time parameter corresponding to the generative adversarial network, and the input data corresponding to the generator is generated according to the random noise, the graph feature data, and the condition information data; The financial fraud sample graph corresponding to the random noise, the graph feature data, and the condition information data is output by the generator.

[0011] Optionally, after the graph structure and the attribute value of the financial fraud sample graph are evaluated according to the preset evaluation index using the discriminator in the generative adversarial network, the financial fraud sample graph meeting the evaluation index is obtained, including: The screening strategy corresponding to the financial fraud sample graph is determined using the discriminator in the generative adversarial network; After the business logic verification and authenticity evaluation of the graph structure and attribute value in the financial fraud sample graph are performed by using the screening strategy, an evaluation quantitative result corresponding to the financial fraud sample graph is obtained. An evaluation quantitative result satisfying a preset evaluation index is obtained, and a financial fraud sample graph corresponding to the current evaluation quantitative result is obtained.

[0012] Optionally, the step of obtaining the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and constructing the financial data set used for training the financial fraud detection model by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph comprises the following steps: The financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph are determined, and a real financial fraud sample graph corresponding to the financial transaction graph and the financial fraud knowledge graph is obtained based on the financial transaction graph. A virtual financial fraud sample graph is determined by using the financial fraud sample graph, and a financial data set used for training the financial fraud detection model is constructed based on the virtual financial fraud sample graph and the real financial fraud sample graph.

[0013] In a second aspect, the present application provides a financial data set determination system, which is used for the training process of a financial fraud detection model, and the system comprises: A knowledge graph construction unit is configured to obtain a financial transaction graph corresponding to original financial transaction data, and construct a financial fraud knowledge graph by using heterogeneous multi-modal features in the original financial transaction data. A generative adversarial network construction unit is configured to determine a virtual fraud transaction graph corresponding to the original financial transaction data according to the financial transaction graph and the financial fraud knowledge graph, construct a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph, and construct a generative adversarial network corresponding to the original financial transaction data by using the generator and the discriminator. A financial fraud sample graph determination unit is configured to determine graph feature data corresponding to the financial transaction graph and condition information data corresponding to the financial fraud knowledge graph, determine a financial fraud sample graph corresponding to the graph feature data and the condition information data by using the generator in the generative adversarial network, and obtain a financial fraud sample graph satisfying a preset evaluation index after the graph structure and attribute value of the financial fraud sample graph are evaluated according to the evaluation index by using the discriminator in the generative adversarial network. A financial data set construction unit is configured to obtain the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and construct a financial data set used for training the financial fraud detection model by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

[0014] In a third aspect, the embodiments of the present application further provide a server, comprising a processor and a memory, the memory storing computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the steps of the financial data set determination method provided in the first aspect.

[0015] In a fourth aspect, the embodiments of the present application further provide a storage medium, which stores computer executable instructions, and the computer executable instructions, when invoked and executed by a processor, cause the processor to implement the steps of the financial data set determination method provided in the first aspect.

[0016] The financial data set determination method, system and server provided by the embodiments of the present application, in the process of constructing a financial data set for training a financial fraud detection model, first acquire a financial transaction graph corresponding to original financial transaction data, and construct a financial fraud knowledge graph using heterogeneous multi-modal features in the original financial transaction data; then determine a virtual fraud transaction graph corresponding to the original financial transaction data according to the financial transaction graph and the financial fraud knowledge graph; construct a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph, and construct a generative adversarial network using the generator and the discriminator; subsequently determine graph feature data corresponding to the financial transaction graph and condition information data corresponding to the financial fraud knowledge graph, determine a financial fraud sample graph corresponding to the graph feature data and the condition information data using the generator in the generative adversarial network, and after the generator in the generative adversarial network evaluates the graph structure and attribute values of the financial fraud sample graph according to a preset evaluation index, obtain the financial fraud sample graph satisfying the evaluation index; finally, obtain a financial transaction graph and a financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and construct a financial data set used for training the financial fraud detection model using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph. This scheme can integrate the graph neural network into the generator and the discriminator of the generative adversarial network, can directly learn and generate a virtual fraud graph sample with complex topological structure, multi-entity association and dynamic evolution characteristics on the graph structure data, so as to obtain a high-quality financial data set for training the financial fraud detection model, thereby significantly improving the performance of the financial fraud detection model.

[0017] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and achieved by the structure particularly pointed out in the description, claims and drawings.

[0018] In order to make the above-mentioned objects, features and advantages of the present application more apparent, the following preferred embodiments are specifically described with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings required to be used in the specific embodiments or prior art description will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0020] Figure 1 The flow chart of the financial data set determination method provided by the embodiment of the present application; Figure 2 The flow chart of step S101 of the financial data set determination method provided by the embodiment of the present application; Figure 3 The flow chart of step S102 of the financial data set determination method provided by the embodiment of the present application, according to the financial transaction graph and the financial fraud knowledge graph, the flow chart of determining the virtual fraud transaction graph corresponding to the original financial transaction data; Figure 4 The flow chart of step S102 of the financial data set determination method provided by the embodiment of the present application, based on the virtual fraud transaction graph, the flow chart of constructing the generator and the discriminator corresponding to the original financial transaction data; Figure 5 The flow chart of step S102 of the financial data set determination method provided by the embodiment of the present application, the flow chart of constructing the generative adversarial network corresponding to the original financial transaction data by using the generator and the discriminator; Figure 6 The flow chart of step S103 of the financial data set determination method provided by the embodiment of the present application, the flow chart of determining the graph feature data corresponding to the financial transaction graph and the condition information data corresponding to the financial fraud knowledge graph, and determining the financial fraud sample graph corresponding to the graph feature data and the condition information data by using the generator in the generative adversarial network; Figure 7 The flow chart of step S103 of the financial data set determination method provided by the embodiment of the present application, after the graph structure and attribute value of the financial fraud sample graph are evaluated according to the preset evaluation index by using the discriminator in the generative adversarial network, the flow chart of obtaining the financial fraud sample graph satisfying the evaluation index; Figure 8 The flow chart of step S104 of the financial data set determination method provided by the embodiment of the present application; Figure 9 The structural schematic diagram of the financial data set determination system provided by the embodiment of the present application; Figure 10 The structural schematic diagram of the server provided by the embodiment of the present application.

[0021] icon: 910 - knowledge graph construction unit; 920 - generative adversarial network construction unit; 930 - financial fraud sample graph determination unit; 940 - financial data set construction unit; 101 - processor; 102 - memory; 103 - bus; 104 - communication interface. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the present application will be described clearly and completely below in conjunction with embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0023] In the fields of finance, insurance, enterprise internal control, etc., using machine learning models to automatically detect financial fraud behaviors (such as false reimbursement, credit card fraud, money laundering, insurance fraud, etc.) has become the main implementation means at present. These fraud detection models are usually supervised learning classifiers, which need a large amount of labeled data for training, including a large number of normal transaction samples and sufficient and diverse fraud transaction samples. Only by learning enough fraud patterns can the model accurately identify new and unknown fraud behaviors in actual application.

[0024] When building a fraud detection model, the core problem faced is the high imbalance of data. In real-world data sets, the number of normal transaction samples is often thousands or even tens of thousands of times that of fraud samples, and fraud samples are not only few in number but also single in pattern. Traditional oversampling methods generate new samples by linear interpolation between existing minority class samples, but this method is prone to generate ambiguous and meaningless samples, and cannot create truly novel fraud patterns, which is of limited help to improve the generalization ability of the model.

[0025] Specifically, the prior art cannot effectively generate fraud samples that are both real and credible and have diverse patterns. Models trained directly using unbalanced data will be severely biased towards the majority class (normal transactions), with a very low recall rate for the minority class (fraudulent transactions), i.e. "better safe than sorry", losing the significance of fraud detection. Simple oversampling techniques (such as SMOTE) generate data of low quality and cannot simulate the complexity and evolution of fraud means in the real world.

[0026] As can be seen from the above, in the training process of the financial fraud detection model, the prior art has the technical problem of insufficient model training, poor generalization ability and low recognition rate of new fraud means due to the extreme scarcity and insufficient diversity of fraud samples.

[0027] Based on this, the present application provides a financial data set determination method, system and server, which can integrate graph neural network into the generator and discriminator of the generative adversarial network, can directly learn and generate virtual fraud graph samples with complex topological structure, multi-entity association and dynamic evolution characteristics on the graph structure data, so as to obtain high-quality financial data set for training of the financial fraud detection model, thereby significantly improving the performance of the financial fraud detection model.

[0028] In order to facilitate the understanding of the present embodiment, first, a kind of financial data set determination method disclosed in the present application embodiment is introduced in detail, as shown in Figure 1 The financial data set is used in the training process of the financial fraud detection model, and the method comprises: Step S101, the financial transaction graph corresponding to the original financial transaction data is obtained, and the heterogeneous multi-modal features in the original financial transaction data are used to construct a financial fraud knowledge graph.

[0029] This step is mainly used to realize the construction process of the financial transaction graph and the knowledge graph, and specifically, the "financial transaction graph" can be generated based on the original financial transaction data, so as to realize that the node represents the transaction subject (such as enterprise, account), and the edge and attribute represent the transaction relationship (such as transfer amount, time);At the same time, the fraud-related rules are extracted from the heterogeneous multi-modal features (such as transaction note text, amount of money, historical violation record, etc.) of the original data, and a "financial fraud knowledge graph" (containing fraud entities, association rules, etc. Prior knowledge) is constructed.

[0030] Step S102, according to the financial transaction graph and the financial fraud knowledge graph, determine the virtual fraud transaction graph corresponding to the original financial transaction data;Based on the virtual fraud transaction graph, the generator and the discriminator corresponding to the original financial transaction data are constructed, and the generator and the discriminator are used to construct the generative adversarial network corresponding to the original financial transaction data.

[0031] This step is mainly used to construct a unique generative adversarial network, which combines the real transaction structure of the financial transaction graph and the fraud mode in the financial fraud knowledge graph to generate a "virtual fraud transaction graph" (a graph structure simulating fraud scenarios);Based on this virtual graph, a "generator" (responsible for generating fraud samples) and a "discriminator" (responsible for judging the authenticity of the sample) are designed, and the two are combined to form a generative adversarial network, which provides tool support for subsequent sample generation.

[0032] In step S103, the graph feature data corresponding to the financial transaction graph and the condition information data corresponding to the financial fraud knowledge graph are determined, the generator in the generative adversarial network is used to determine the financial fraud sample graph corresponding to the graph feature data and the condition information data, and the discriminator in the generative adversarial network is used to evaluate the graph structure and attribute value of the financial fraud sample graph according to a preset evaluation index, so that the financial fraud sample graph satisfying the evaluation index is obtained.

[0033] This step is mainly used for generating and screening effective financial fraud sample graphs. The graph feature data (such as node connection rule) of the financial transaction graph and the condition information data (such as fraud rule constraint) of the financial fraud knowledge graph can be extracted and input into the generator of the generative adversarial network. After the generator generates the financial fraud sample graph in combination with the graph feature data and the condition information data, the discriminator evaluates the sample graph according to the preset evaluation index (such as graph structure rationality and attribute value authenticity), and finally screens out the effective financial fraud sample graph satisfying the index.

[0034] In step S104, the financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph are obtained based on the financial transaction graph, and the financial data set used for training the financial fraud detection model is constructed by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

[0035] This step takes the original financial transaction graph as a reference to determine the basic transaction information and knowledge association corresponding to the financial fraud sample graph. Finally, the effective financial fraud sample graph (fraud sample), the original financial transaction graph (normal sample) and the financial fraud knowledge graph (auxiliary knowledge) are integrated to form the financial data set required for training the financial fraud detection model.

[0036] Optionally, the step S101 of obtaining the financial transaction graph corresponding to the original financial transaction data and constructing the financial fraud knowledge graph by using the heterogeneous multi-modal features in the original financial transaction data, as shown in FIG. 1, includes the following steps. Figure 2 In step S201, the data source corresponding to the financial data set is obtained, and the original financial transaction data is determined by using the enterprise transaction data, the bank flow data, the accounting voucher data, the enterprise business data, the judicial judgment data and the media report data determined by the data source.

[0037] The collection process of the original financial transaction data can be obtained from the corresponding data source, for example, the original financial transaction data including normal transaction data and a small amount of labeled real fraud transaction data is collected from the enterprise internal transaction system, bank flow, accounting voucher, enterprise business information, judicial judgment book and public media report. These data usually include transaction subject, transaction object, transaction amount, time, place, transaction description, account information and the like.

[0038] ​In step S202, the original financial transaction data is standardized and encoded according to the data features and category features corresponding to the original financial transaction data, and the heterogeneous feature data corresponding to the original financial transaction data is obtained.

[0039] After obtaining the original financial transaction data, the heterogeneous feature extraction and encoding processing are performed on the original financial transaction data, and the original financial transaction data is cleaned, the missing value is processed, and the normalization is performed. For numerical features (such as transaction amount, frequency, time interval), normalization or standardization processing is performed; for category features (such as transaction type, account type, business scenario), one-hot encoding or entity embedding is used for encoding. In actual scenarios, special attention can be paid to the fusion of multi-modal features, for example, semantic features are extracted from transaction description text (such as using BERT / Transformer model), and structured information is extracted from scanned bill images.

[0040] In step S203, the specific entity data and the specific relationship data corresponding to the original financial transaction data are determined by using the heterogeneous feature data, the nodes corresponding to the financial transaction graph are determined according to the specific entity data, and the edges corresponding to the financial transaction graph are determined according to the specific relationship data, and then the financial transaction graph is constructed by using the nodes and the edges.

[0041] After collecting the original financial transaction data and extracting the multi-modal heterogeneous feature data therefrom, the original financial transaction data is constructed into a financial transaction graph, which is used to capture the multi-hop association and the fund flow path in the transaction. In the construction process of the financial transaction graph (FTG), the original financial transaction data can be constructed into a graph structure: specific entities, such as accounts, users, and enterprises, are used as nodes, and specific relationships, such as transactions, transfers, associations, and common addresses, are used as edges, and the attributes of the edges can include transaction amount, timestamp, etc.

[0042] The FTG is used as the input of the subsequent graph neural network module to capture the multi-hop association, community structure, and fund flow path implied in the transaction data, and these complex associations are the key to financial fraud identification.

[0043] In step S204, the key entity data, the key relationship data, and the financial fraud pattern data corresponding to the original financial transaction data are determined by using the heterogeneous feature data, the embedding representation vector corresponding to the financial fraud knowledge graph is determined by using the key entity data, the key relationship data, and the financial fraud pattern data, and the financial fraud knowledge graph is constructed according to the semantic information and the structural information of the embedding representation vector.

[0044] After collecting the original financial transaction data and extracting the multi-modal heterogeneous feature data therefrom, a financial fraud knowledge graph is constructed based on the data, specifically to encode entities, relationships and known fraud patterns, and generate embedded representations thereof as conditional inputs of a subsequent generative adversarial network. Meanwhile, in the process of constructing the financial fraud knowledge graph (FKG), entity recognition and relationship extraction are specifically performed to identify and extract key entities (such as individuals, enterprises, accounts, transactions, IP addresses, device IDs, etc.) and their relationships (such as “transfer to”, “is the legal representative”, “joint investment”, “registered in”, etc.) from structured and unstructured original financial transaction data.

[0045] The pattern encoding and embedding process mainly encodes known financial fraud patterns (for example, common means of internal fraud, money laundering paths, and characteristics of related transactions) as specific nodes and relationship patterns in the knowledge graph. Related knowledge graph embedding technologies, such as TransE, RotatE, ComplEx or GraphSAGE, are used to represent entities and relationships in the knowledge graph as low-dimensional vectors (embeddings) to capture complex semantic and structural information.

[0046] The FKG serves as a conditional input of the generative adversarial network to guide the generator to generate virtual fraud samples that are more consistent with real business logic and fraud pattern rules, thereby addressing the shortcomings of traditional generative adversarial networks in generating samples with complex association relationships and business rule constraints.

[0047] Optionally, the virtual fraud transaction graph corresponding to the original financial transaction data is determined according to the financial transaction graph and the financial fraud knowledge graph, as shown in Figure 3 , which includes: Step S301, constructing a graph neural network corresponding to the original financial transaction data using the financial transaction graph and the financial fraud knowledge graph; Step S302, obtaining time sequence dependency parameters, association parameters, heterogeneous node feature parameters and heterogeneous edge feature parameters corresponding to the financial transaction graph and the financial fraud knowledge graph based on the graph neural network; Step S303, generating a virtual fraud transaction graph according to the time sequence dependency parameters, the association parameters, the heterogeneous node feature parameters and the heterogeneous edge feature parameters.

[0048] In a specific scenario, a unique conditional generative adversarial network (KG-GNN-CGAN) is finally constructed by using the financial transaction graph and the financial fraud knowledge graph, which combines the advantages of knowledge graph and graph neural network, and enhances the financial fraud samples in a conditional generation manner. The graph neural network enables it to learn and generate virtual financial fraud graph samples with complex correlation structure, temporal dependence and multi-modal heterogeneous features, so as to obtain the temporal dependence parameters, correlation parameters, heterogeneous node feature parameters and heterogeneous edge feature parameters corresponding to the financial transaction graph and the financial fraud knowledge graph based on the graph neural network, and generate virtual fraud transaction graph by using the temporal dependence parameters, correlation parameters, heterogeneous node feature parameters and heterogeneous edge feature parameters.

[0049] Optionally, the generator and the discriminator corresponding to the original financial transaction data are constructed based on the virtual fraud transaction graph, as shown in Figure 4 as shown, comprising: Step S401, determining the financial fraud knowledge graph corresponding to the graph neural network based on the virtual fraud transaction graph, and obtaining the first input data and the first output data corresponding to the financial fraud knowledge graph; Step S402, determining the financial transaction graph corresponding to the graph neural network based on the virtual fraud transaction graph, and obtaining the second input data and the second output data corresponding to the financial transaction graph; Step S403, constructing the generator by using the first input data and the first output data, and constructing the discriminator by using the second input data and the second output data.

[0050] In the construction process of the generator (Generator, G), the network structure corresponding to the generator can be based on the encoder-decoder structure of the graph neural network (GNN, such as GCN / GraphSAGE / GAT). The input data is the random noise vector z and the conditional input c_KG (for example, the knowledge graph embedding vector of a specific fraud mode or the subgraph structure generated by the knowledge graph) from the knowledge graph, and the graph structure feature c_GNN (for example, the node / edge embedding of the real fraud subgraph extracted by the GNN encoder) from the real financial transaction graph.

[0051] The generator can integrate multi-head attention mechanism or Transformer encoder, so as to effectively process and generate virtual fraud transaction graph with temporal dependence, multi-hop correlation and heterogeneous node / edge features.

[0052] The output result contains virtual financial fraud sample graph G(z, c_KG, c_GNN) with similar distribution as the real complex correlation financial fraud sample. These virtual samples contain nodes (such as accounts, entities) and edges (such as transactions), as well as their attributes (such as amount, time, description).

[0053] In the construction of the discriminator (D), its network structure is also based on the discriminator of a graph neural network (GNN). The input data are real financial fraud sample image x_real and virtual financial fraud sample image x_fake = G(z, c_KG, c_GNN) generated by the generator.

[0054] During graph feature extraction, the discriminator utilizes specific GNN layers, such as multi-layer graph convolutional networks, to extract high-order graph structure features and node / edge embeddings from the input graph structure. The final output is a scalar representing the probability that the input graph is a genuine fraudulent sample. Specifically, the discriminator, based on a graph neural network, is responsible for distinguishing between genuine fraudulent graphs and generated virtual fraudulent graphs. It undergoes adversarial training by incorporating an optimization objective based on graph structure loss to ensure the authenticity of the graph structure and attributes of the generated samples.

[0055] Optionally, a generative adversarial network corresponding to the original financial transaction data can be constructed using a generator and a discriminator, such as... Figure 5 As shown, it includes: Step S501: Use the generator to generate a virtual fraud diagram corresponding to the original financial transaction data; Step S502: Obtain the real fraud diagram contained in the original financial transaction data, and use the discriminator to obtain the discrimination result between the real fraud diagram and the virtual fraud diagram; Step S503: Train the generator and discriminator alternately using the WGAN-GP loss function or the Wasserstein distance loss function, and obtain the discrimination results in real time; Step S504: When the discrimination result meets the preset threshold condition, construct a generative adversarial network based on the current generator and discriminator.

[0056] The above process describes the training of a generative adversarial network using a loss function. The objective function can be a GAN objective function, such as the WGAN-GP loss function or an improved Wasserstein distance loss function, combined with specific graph-related losses such as graph structure loss (adjacency matrix reconstruction loss) or node / edge attribute loss, to stabilize training and improve the graph structure and attribute authenticity of the generated samples.

[0057] During training, the generator and discriminator need to be trained alternately. In the specific training process, the discriminator is committed to distinguishing between real and virtual fraudulent graphs, while the generator tries to generate fraudulent graph samples that can "deceive" the discriminator and satisfy the knowledge graph constraints and graph structure features.

[0058] Optionally, the graph feature data corresponding to the financial transaction graph and the conditional information data corresponding to the financial fraud knowledge graph are determined. Then, a generator in a generative adversarial network is used to determine the financial fraud sample graphs corresponding to the graph feature data and the conditional information data, such as...Figure 6 As shown in the figure, comprising: Step S601, determining the extraction strategy corresponding to the financial fraud sample graph based on the original financial transaction data; Step S602, extracting and obtaining graph feature data from the financial transaction graph using the extraction strategy, and extracting and obtaining condition information data from the financial fraud knowledge graph using the extraction strategy; Step S603, setting random noise using the time parameter corresponding to the generative adversarial network, and generating input data corresponding to the generator according to the random noise, the graph feature data and the condition information data; Step S604, controlling the generator to output the financial fraud sample graph corresponding to the random noise, the graph feature data and the condition information data.

[0059] The financial fraud sample graph is essentially a virtual fraud sample. After the generative adversarial network is trained and converged, the corresponding generator is fixed. The random noise, the condition information c_KG extracted from the knowledge graph according to the real fraud mode, and the graph feature c_GNN extracted from a small number of real complex associated fraud cases are inputted, so as to generate a large number of financial fraud sample graphs with virtual complex associations in batches.

[0060] It is worth mentioning that the generated sample is graph structure data, which contains multiple interrelated nodes and edges, and can simulate complex patterns such as "gang fraud", "money laundering chain" and "false associated transaction".

[0061] Optionally, after the graph structure and attribute value of the financial fraud sample graph are evaluated according to the preset evaluation index by using the discriminator in the generative adversarial network, the financial fraud sample graph satisfying the evaluation index is obtained, such as Figure 7 As shown in the figure, comprising: Step S701, determining the screening strategy corresponding to the financial fraud sample graph by using the discriminator in the generative adversarial network; Step S702, after the graph structure and attribute value in the financial fraud sample graph are subjected to business logic verification and authenticity evaluation by using the screening strategy, obtaining the evaluation quantization result corresponding to the financial fraud sample graph; Step S703, obtaining the evaluation quantization result satisfying the preset evaluation index, and obtaining the financial fraud sample graph corresponding to the current evaluation quantization result.

[0062] The generated sample is subjected to business logic verification and authenticity evaluation to ensure its high quality and high fidelity. The business logic verification mainly performs preliminary screening on the generated virtual sample, and uses business rules or knowledge graph constraints to eliminate samples that do not obviously comply with the financial business logic or have too low quality. The authenticity evaluation process can introduce specific evaluation indexes, such as graph isomorphism index, node / edge attribute distribution similarity, to quantitatively evaluate the authenticity of the graph structure and attributes of the generated virtual sample.

[0063] Optionally, the step S104 of obtaining the financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph based on the financial transaction graph, and constructing the financial data set used for training the financial fraud detection model by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph, as shown in Figure 8 Step S801, determine the financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph, and obtain the real financial fraud sample graph corresponding to the financial transaction graph and the financial fraud knowledge graph based on the financial transaction graph; Step S802, determine the virtual financial fraud sample graph by using the financial fraud sample graph, and construct the financial data set used for training the financial fraud detection model based on the virtual financial fraud sample graph and the real financial fraud sample graph.

[0064] After obtaining the financial fraud sample graph, the virtual financial fraud sample graph is determined by using the financial fraud sample graph. In addition, the financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph are obtained, and the real financial fraud sample graph corresponding to the financial transaction graph and the financial fraud knowledge graph is obtained based on the financial transaction graph. The virtual financial fraud sample graph and the real financial fraud sample graph are integrated to form a financial data set with larger scale, more abundant mode and more balanced positive and negative sample distribution.

[0065] The embodiment of the application innovatively integrates the graph neural network GNN into the generator and discriminator of the generative adversarial network CGAN. Through the graph convolution, graph attention mechanism and other capabilities of GNN, virtual fraud graph samples with complex topological structure (such as star, chain, gang network), multi-entity association and dynamic evolution characteristics can be directly learned and generated on the graph structure data. This graph-based generation method not only generates samples that contain clear fraud relationship network, but also significantly improves the authenticity of the generated samples and the simulation diversity of complex associated fraud patterns, far beyond the limitations of the prior art in simulating such high-order complex patterns.

[0066] The embodiment of the application constructs the financial fraud knowledge graph (FKG), and embeds the entities and relationships of the FKG and the known fraud mode code as the conditional input of the CGAN. This conditional guidance mechanism enables the generator to strictly follow the business rules and fraud logic (for example, legal paths of fund flow, identity relationships of associated entities, typical characteristics of specific fraud activities, etc.) encoded in the knowledge graph when generating virtual samples. Therefore, the generated virtual fraud samples not only have real data distribution, but also achieve an unprecedented level of accuracy in business logic and fraud patterns, effectively avoiding the generation of "meaningless" false samples and ensuring the high-quality contribution of enhanced data to downstream fraud identification models.​

[0067] Furthermore, by integrating the precise guidance of the generation process provided by knowledge graphs with the powerful generation capabilities of graph neural networks for complex association patterns, this invention can efficiently and purposefully generate a large number of virtual samples that conform to specific complex association rules and can simulate long-tail fraud patterns. These high-quality synthetic samples effectively expand the scarce real fraud dataset, significantly improve the imbalance between positive and negative samples, and enable downstream fraud detection models to fully learn these complex and rare fraud features. This significantly improves the detection accuracy and recall rate for novel, covert, and long-tail complex association financial frauds, effectively reducing the risk of false negatives.

[0068] The financial datasets obtained using the financial dataset determination method in this embodiment of the invention generate virtual samples that are not only numerous but, more importantly, far surpass existing technologies in terms of pattern diversity, structural complexity, and logical rationality. By using these more comprehensive and representative augmented data to train downstream financial fraud detection models, the models can learn broader and more detailed fraud features and patterns. This enables the financial fraud detection models to have stronger generalization capabilities for unknown fraud patterns and greater robustness against adversarial fraud, thus providing financial institutions and enterprises with a more efficient and reliable intelligent anti-fraud early warning method.

[0069] As can be seen from the financial dataset determination method in the above embodiments, this method can integrate graph neural networks into the generator and discriminator of generative adversarial networks, and can directly learn and generate virtual fraud graph samples with complex topological structures, multi-entity associations and dynamic evolution characteristics on graph structure data, thereby obtaining high-quality financial datasets for training financial fraud detection models, thus significantly improving the performance of financial fraud detection models.

[0070] Corresponding to the above-described method for determining financial datasets, this embodiment of the invention also provides a system for determining financial datasets, which are used in the training process of a financial fraud detection model, such as... Figure 9 As shown, the system includes: The knowledge graph construction unit 910 is used to obtain the financial transaction graph corresponding to the original financial transaction data and to construct a financial fraud knowledge graph using the heterogeneous multimodal features in the original financial transaction data. Generative adversarial network building unit 920 is used to determine the virtual fraud transaction graph corresponding to the original financial transaction data based on the financial transaction graph and the financial fraud knowledge graph; to build a generator and discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph; and to build a generative adversarial network corresponding to the original financial transaction data using the generator and discriminator. The financial fraud sample graph determination unit 930 is configured to determine graph feature data corresponding to the financial transaction graph and condition information data corresponding to the financial fraud knowledge graph, determine a financial fraud sample graph corresponding to the graph feature data and the condition information data by using a generator in the generative adversarial network, and obtain the financial fraud sample graph satisfying the evaluation index by evaluating the graph structure and the attribute value of the financial fraud sample graph according to a preset evaluation index by using a discriminator in the generative adversarial network. The financial data set construction unit 940 is configured to obtain the financial transaction graph corresponding to the financial fraud sample graph and the financial fraud knowledge graph based on the financial transaction graph, and construct a financial data set used for training of the financial fraud detection model by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

[0071] It can be known from the above financial data set determination system that the system can integrate the graph neural network into the generator and the discriminator of the generative adversarial network, can directly learn and generate a virtual fraud graph sample with complex topological structure, multi-entity association and dynamic evolution characteristics on the graph structure data, thereby obtaining a high-quality financial data set for training of the financial fraud detection model, and significantly improving the performance of the financial fraud detection model.

[0072] The financial data set determination system provided in the embodiment of the application has the same implementation principle and technical effects as the financial data set determination method, and for brevity of description, the part of the system embodiment not mentioned can be referred to the corresponding content in the financial data set determination method.

[0073] The embodiment also provides a server, and a structure diagram of the server is shown in Figure 10 The device includes a processor 101 and a memory 102; the memory 102 is configured to store one or more computer instructions, and the one or more computer instructions are executed by the processor to implement the steps of the financial data set determination method.

[0074] Figure 10 The server shown in the figure also includes a bus 103 and a communication interface 104, and the processor 101, the communication interface 104 and the memory 102 are connected through the bus 103.

[0075] The memory 102 can include a high-speed random access memory (RAM) and can also include a non-volatile memory, for example, at least one disk memory. The bus 103 can be an ISA bus, a PCI bus or an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, Figure 10Only one bidirectional arrow is used to represent multiple buses or multiple types of buses.

[0076] The communication interface 104 is configured to connect with at least one user terminal and other network units through a network interface, and transmit the encapsulated IPv4 packet or IPv4 packet to the user terminal through the network interface.

[0077] The processor 101 can be an integrated circuit chip with processing capability. In implementation process, each step of the above method can be completed by integrated logic circuit or software form instruction in the processor 101. The processor 101 described above can be a general processor, including a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. Each method, step and logic block in the embodiments of the present disclosure can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present disclosure can be directly embodied as a hardware code processor to execute, or be executed by a combination of hardware and software modules in the code processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium in the art. The storage medium is located in the memory 102, and the processor 101 reads the information in the memory 102, and combines the hardware to complete the steps of the method of the above embodiment.

[0078] The embodiment of the present application further provides a storage medium, and the storage medium stores a computer program. When the computer program is run by a processor, the steps of the financial data set determination method in the foregoing embodiment are executed.

[0079] In several embodiments provided in the present application, it should be understood that the disclosed system, device, apparatus and method can be implemented in other manners. The above described system embodiments are merely illustrative. For example, the division of the units is only a logical function division. There can be another division manner for the actual implementation, or a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, or the among different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0080] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments of the present application.

[0081] In addition, each functional unit in the various embodiments of the present application can be integrated in one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit.

[0082] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0083] Finally, it should be noted that the above-described embodiments are merely specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit the same. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that any person skilled in the art can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to some of the technical features, within the technical scope disclosed by the present application. The modifications, changes or replacements do not cause the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A financial dataset determination method characterized by, The financial data set is used for a training process of a financial fraud detection model, and the method comprises: acquiring a financial transaction graph corresponding to original financial transaction data, and constructing a financial fraud knowledge graph by using heterogeneous multi-modal features in the original financial transaction data; determining a virtual fraud transaction graph corresponding to the original financial transaction data according to the financial transaction graph and the financial fraud knowledge graph, constructing a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph, and constructing a generative adversarial network corresponding to the original financial transaction data by using the generator and the discriminator; determining graph feature data corresponding to the financial transaction graph and conditional information data corresponding to the financial fraud knowledge graph, determining a financial fraud sample graph corresponding to the graph feature data and the conditional information data by using the generator in the generative adversarial network, and obtaining the financial fraud sample graph satisfying a preset evaluation index after evaluating a graph structure and an attribute value of the financial fraud sample graph by using the discriminator in the generative adversarial network according to the evaluation index; acquiring the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and constructing the financial data set used for training the financial fraud detection model by using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

2. The financial dataset determination method of claim 1, wherein, The step of acquiring a financial transaction graph corresponding to original financial transaction data, and constructing a financial fraud knowledge graph by using heterogeneous multi-modal features in the original financial transaction data, comprises: acquiring a data source corresponding to the financial data set, and determining the original financial transaction data by using enterprise transaction data, bank flow data, accounting voucher data, enterprise business data, judicial decision data and media report data determined by using the data source; after standardizing and encoding the original financial transaction data according to data features and category features corresponding to the original financial transaction data, obtaining heterogeneous feature data corresponding to the original financial transaction data; determining specific entity data and specific relationship data corresponding to the original financial transaction data by using the heterogeneous feature data, determining nodes corresponding to the financial transaction graph according to the specific entity data, and determining edges corresponding to the financial transaction graph according to the specific relationship data, and then constructing the financial transaction graph by using the nodes and the edges; determining key entity data, key relationship data and financial fraud mode data corresponding to the original financial transaction data by using the heterogeneous feature data, determining embedding representation vectors corresponding to the financial fraud knowledge graph by using the key entity data, the key relationship data and the financial fraud mode data, and constructing the financial fraud knowledge graph according to semantic information and structural information corresponding to the embedding representation vectors.

3. The financial dataset determination method of claim 1, wherein, The step of determining a virtual fraud transaction graph corresponding to the original financial transaction data according to the financial transaction graph and the financial fraud knowledge graph, comprises: construct a graph neural network corresponding to the original financial transaction data by using the financial transaction graph and the financial fraud knowledge graph; obtain time sequence dependency parameters, relevance parameters, heterogeneous node feature parameters and heterogeneous edge feature parameters corresponding to the financial transaction graph and the financial fraud knowledge graph based on the graph neural network; generate the virtual fraud transaction graph according to the time sequence dependency parameters, the relevance parameters, the heterogeneous node feature parameters and the heterogeneous edge feature parameters.

4. The financial dataset determination method of claim 3, wherein, construct a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph, including: determine the financial fraud knowledge graph corresponding to the graph neural network based on the virtual fraud transaction graph, and obtain first input data and first output data corresponding to the financial fraud knowledge graph; determine the financial transaction graph corresponding to the graph neural network based on the virtual fraud transaction graph, and obtain second input data and second output data corresponding to the financial transaction graph; construct the generator by using the first input data and the first output data, and construct the discriminator by using the second input data and the second output data.

5. The financial dataset determination method of claim 3, wherein, construct a generative adversarial network corresponding to the original financial transaction data by using the generator and the discriminator, including: generate a virtual fraud graph corresponding to the original financial transaction data by using the generator; obtain a real fraud graph contained in the original financial transaction data, and obtain a discrimination result of the real fraud graph and the virtual fraud graph by using the discriminator; alternately train the generator and the discriminator by using a WGAN-GP loss function or a Wasserstein distance loss function, and obtain the discrimination result in real time; when the discrimination result meets a preset threshold condition, construct the generative adversarial network based on the current generator and the discriminator.

6. The financial dataset determination method of claim 1, wherein, determine graph feature data corresponding to the financial transaction graph and condition information data corresponding to the financial fraud knowledge graph, and determine a financial fraud sample graph corresponding to the graph feature data and the condition information data by using the generator in the generative adversarial network, including: determine an extraction strategy corresponding to the financial fraud sample graph based on the original financial transaction data; extract the graph feature data from the financial transaction graph by using the extraction strategy, and extract the condition information data from the financial fraud knowledge graph by using the extraction strategy; set random noise by using a time parameter corresponding to the generative adversarial network, and generate input data corresponding to the generator according to the random noise, the graph feature data and the condition information data; control the generator to output the financial fraud sample graph corresponding to the random noise, the graph feature data and the condition information data.

7. The financial dataset determination method of claim 1, wherein, after evaluating the graph structure and attribute value of the financial fraud sample graph according to a preset evaluation index by using the discriminator in the generative adversarial network, obtain the financial fraud sample graph meeting the evaluation index, including: determine a screening strategy corresponding to the financial fraud sample graph by using the discriminator in the generative adversarial network; After using the screening strategy to perform business logic verification and authenticity assessment on the graph structure and attribute values ​​in the financial fraud sample graph, the evaluation quantification result corresponding to the financial fraud sample graph is obtained. Obtain the quantitative evaluation result that meets the preset evaluation indicators, and obtain the financial fraud sample image corresponding to the current quantitative evaluation result.

8. The financial dataset determination method of claim 1, wherein, The steps of obtaining the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and constructing the financial dataset used for training the financial fraud detection model using the financial fraud sample graph, the financial transaction graph, and the financial fraud knowledge graph, include: Determine the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph, and obtain the real financial fraud sample graph corresponding to the financial transaction graph and the financial fraud knowledge graph based on the financial transaction graph; The virtual financial fraud sample map is determined using the financial fraud sample map, and the financial dataset used for training the financial fraud detection model is constructed based on the virtual financial fraud sample map and the real financial fraud sample map.

9. A financial dataset determination system, characterized by, The financial dataset is used for training the financial fraud detection model, and the system includes: The knowledge graph construction unit is used to obtain the financial transaction graph corresponding to the original financial transaction data, and to construct a financial fraud knowledge graph using the heterogeneous multimodal features in the original financial transaction data. A generative adversarial network (GAN) construction unit is used to determine a virtual fraud transaction graph corresponding to the original financial transaction data based on the financial transaction graph and the financial fraud knowledge graph; construct a generator and a discriminator corresponding to the original financial transaction data based on the virtual fraud transaction graph; and construct a generative adversarial network corresponding to the original financial transaction data using the generator and the discriminator. The financial fraud sample graph determination unit is used to determine the graph feature data corresponding to the financial transaction graph and the condition information data corresponding to the financial fraud knowledge graph. It uses the generator in the generative adversarial network to determine the financial fraud sample graph corresponding to the graph feature data and the condition information data, and uses the discriminator in the generative adversarial network to evaluate the graph structure and attribute values ​​of the financial fraud sample graph according to the preset evaluation index, and then obtains the financial fraud sample graph that meets the evaluation index. The financial dataset construction unit is used to obtain the financial transaction graph and the financial fraud knowledge graph corresponding to the financial fraud sample graph based on the financial transaction graph, and to construct the financial dataset used for training the financial fraud detection model using the financial fraud sample graph, the financial transaction graph and the financial fraud knowledge graph.

10. A server, characterized by The device includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the steps of the financial dataset determination method according to any one of claims 1 to 8.

Images